Automatically growing global reactive neural network potential energy surfaces: A trajectory-free active learning...

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Machine Learning

Background:

• Accurate potential energy surfaces (PESs) are crucial for understanding chemical reactions.
• Neural networks (NNs) are powerful tools for constructing PESs but require efficient data sampling.
• Traditional methods often rely on computationally expensive trajectory simulations.

Purpose of the Study:

• To develop an efficient, trajectory-free active learning method for constructing globally accurate reactive PESs using NNs.
• To address the challenge of data point selection for NN-based PESs without predictive variance.
• To accelerate the convergence of quantum scattering calculations.

Main Methods:

• Proposed an active learning strategy that minimizes the negative of the squared difference surface (NSDS) between two NN models.
• Iteratively added data points from NSDS minima to the training set to improve PES accuracy.
• Avoided the need for classical trajectory or molecular dynamics simulations.

Main Results:

• Demonstrated the efficiency and robustness of the proposed method on H3 and OH3 reactive systems.
• Achieved fast convergence of reactive PESs with respect to the number of training points.
• Successfully constructed globally accurate PESs enabling reliable quantum scattering probabilities.

Conclusions:

• The developed trajectory-free active learning method is an efficient approach for building accurate reactive PESs with NNs.
• This strategy significantly reduces the computational cost associated with PES construction.
• The method shows great promise for accelerating theoretical chemical dynamics studies.